Velocity measurements from drifter GPS records are assimilated and used in an ensemble-based inversion technique to extract the river bathymetry. The method is tested on a deep meandering reach and a shallow braided reach of the Kootenai River in Idaho, USA. The Regional Ocean Modeling System (ROMS) is used to model numerous statistically varied bathymetries to create an ensemble of hydrodynamic states. These states, the drifter observations, and the uncertainty of each are combined to form a cost function which is minimized to produce an estimated velocity ﬁeld. State augmentation is then used to relate the velocity ﬁeld to bathymetry. Our goals are to assess whether ROMS can accurately reproduce the Kootenai River ﬂow to an extent that depth inversion is feasible, investigate if drifter paths are sensitive enough to bottom topography to make depth inversion possible, and to establish practical limitations of the present methodology. At both test sites, the depth inversion method produced an estimate of bathymetry that was more accurate and more skillful than the prior estimate.